Here's what we just covered.

A network starts out knowing nothing. Every connection between neurons has a weight, and at the beginning those weights are random.

The network makes a prediction, compares it to the right answer, measures how far off it was, and uses that gap to figure out which weights to nudge, and by how much.

Do that millions of times, across thousands of examples, and the weights slowly settle into a configuration that produces good answers.

Nobody programs what the hidden layers look for. Nobody writes down the rules. The network discovers them on its own, because certain internal patterns turn out to be useful for getting the right answer.

The measure of success is generalization: not getting the training examples right, but getting examples it has never seen right.

That's the whole mechanism. It sounds almost too simple for something that can recognize a face, translate a sentence, or write a paragraph. But this is genuinely how it works.

What comes next is a stranger question. Once training is done, what remains? The network has "learned" something. But where is that knowledge? What form does it take? And in what sense does a collection of numbers actually know anything at all?